From 2 December 2027, a high-risk AI system listed in Annex III must be designed and developed in such a way that it achieves an appropriate level of accuracy, robustness and cybersecurity, and performs consistently in those respects throughout its lifecycle (Article 15(1) AI Act). That requirement rests on the provider. A municipality buying the system does not build its accuracy itself, but it has obligations of its own that depend on it: it uses the system in accordance with the instructions for use, ensures that input data are relevant and sufficiently representative to the extent it controls them, monitors the operation and intervenes when things go wrong (Article 26). And because the system makes decisions about residents, the GDPR already asks it today for accurate data and appropriate security.
Suppose the Work and Income department is considering an AI system that assesses applications for social assistance and indicates for each application whether an additional check is needed. The vendor's bid states "an accuracy of 94 per cent". The buyer wants to know whether that is good, the CISO asks what happens if someone tries to mislead the model, and the privacy officer asks how the municipality corrects wrong outcomes. Article 15 offers a framework for all three questions, but no number.
In short
Article 15 has five paragraphs. Paragraph 1 requires an appropriate level of accuracy, robustness and cybersecurity and consistent performance throughout the lifecycle. Paragraph 2 has the Commission encourage the development of benchmarks and measurement methodologies. Paragraph 3 requires the levels of accuracy and the relevant accuracy metrics to be declared in the instructions for use; Article 13(3)(b)(ii) spells this out for accuracy, robustness and cybersecurity. Paragraph 4 requires resilience against errors and faults, possibly with a fail-safe plan, and requires systems that continue to learn after deployment to eliminate or reduce as far as possible the risk of feedback loops. Paragraph 5 requires resilience against attacks, with measures against data poisoning, model poisoning, adversarial examples, confidentiality attacks and model flaws. A certificate under Regulation (EU) 2019/881 or compliance with the Cyber Resilience Act, Regulation (EU) 2024/2847, can create a presumption of conformity, but only for cybersecurity (Article 42(2) and (3)). For Annex III these requirements apply from 2 December 2027; for systems placed on the market or put into service before that date, the transitional rule of Article 111(2) applies.
This piece covers Article 15 as a whole and what provider and municipality do with it. Techniques for making models explainable, and whether outputs are reproducible, are covered in explainable AI and the black box. How to measure and reduce bias in the data itself is covered in data quality and bias mitigation. The duty of care and the reporting duty of the Dutch Cybersecurity Act are covered in AI agents and the Dutch Cybersecurity Act.
What Article 15 requires, paragraph by paragraph
| Paragraph | What the Regulation requires | What the municipality does with it |
|---|---|---|
| 1 | Appropriate level of accuracy, robustness and cybersecurity, and consistent performance throughout the lifecycle | Asks which level the provider considers appropriate for its purpose, and why |
| 2 | The Commission encourages benchmarks and measurement methodologies, with metrology and benchmarking authorities among others | Asks which measurement method the provider used |
| 3 | Levels of accuracy and the relevant accuracy metrics are declared in the instructions for use | Reads the figures, tests them against its own population and monitors against them |
| 4 | Resilient against errors and faults; redundancy or fail-safe plans possible; feedback loops in systems that keep learning eliminated or reduced | Sets up a fallback process and agrees how the system learns from municipal data |
| 5 | Resilient against attacks; AI-specific measures against data poisoning and adversarial examples, among others | Asks about tests and measures, and secures its own environment around the system |
The requirements of Article 15 are in Chapter III, Section 2. The provider ensures that the system complies with them (Article 16(a)). If the municipality builds the system itself, or has it built and puts it into service under its own name, it is the provider itself (Article 3, point 3). The same can happen if it substantially modifies a high-risk system (Article 25(1)). Article 15 then rests on the municipality.
Paragraph 1: an appropriate level, throughout the lifecycle
The Regulation gives no percentage. "Appropriate" depends on the intended purpose, and recital 74 adds: in line with the generally acknowledged state of the art. A system that flags social assistance applications for additional checks affects people's livelihoods. An error costs a resident time, money or trust. That justifies a higher bar than for a system that sorts internal mail.
The second half of paragraph 1 matters at least as much: performance must be consistent throughout the lifecycle. A model that scores well at delivery can get worse when reality changes, for example after a change in the law, a new target group or a different way of applying. That is called drift. Article 15 asks the provider to design the system against it; the monitoring by the provider (Article 72) and by the municipality (Article 26(5)) must make it visible.
Paragraph 3 and Article 13: the figures are in the instructions for use
Paragraph 3 says that the levels of accuracy and the relevant accuracy metrics are declared in the instructions for use. Article 13(3)(b)(ii) widens this: the instructions contain the level of accuracy, including its metrics, robustness and cybersecurity against which the system has been tested and validated and which can be expected, and any known and foreseeable circumstances that may affect that expected level. Recital 74 asks providers to communicate this information clearly and in an easily understandable way, free of misunderstandings or misleading statements. How to read the instructions as a whole is covered in what Article 13 requires of instructions for use.
Back to the 94 per cent. Suppose that 5 per cent of applications turn out to be unjustified after a closer check. A system that marks every application as "fine" is then 95 per cent accurate and finds not a single problem. One number says little. In our reading, a municipality only gets something useful out of the paragraph 3 figures if they show at least:
- how many applications the system flags wrongly (false positives): residents who get an unnecessary additional check;
- how many unjustified applications it misses (false negatives);
- whether those error rates differ between groups, for example by age, household composition or neighbourhood, and on which test data that was measured;
- at which threshold the system flags an application, and what happens if the municipality moves that threshold;
- under which circumstances the figures no longer hold, such as a target group different from the one in the test data.
That is our practical translation, not a list from the Regulation. It does align with the European Data Protection Board (EDPB), which lists as a design element of the accuracy principle that processing is measurably accurate, with reducing false positives and false negatives in AI and automated decisions as its example (Guidelines 4/2019, point 79).
Paragraph 2: benchmarks, measurement methods and standards
Paragraph 2 does not impose an obligation on providers but a task on the Commission: it encourages, as appropriate, the development of benchmarks and measurement methodologies, together with stakeholders and organisations such as metrology and benchmarking authorities. Recital 74 refers to Union law on legal metrology and to cooperation with international partners.
In practice this means that there is currently no legally prescribed method for measuring the accuracy or robustness of a high-risk system. The provider chooses how to measure and must be able to justify that choice. Harmonised standards may later create a presumption of conformity, but only once their references are published in the Official Journal and only to the extent they cover the requirements (Article 40(1)). At the time of writing, the only harmonised standard under the AI Act we know of is EN 18286:2026 on quality management (Implementing Decision (EU) 2026/2274). We know of no harmonised standard yet for accuracy, robustness or cybersecurity. A vendor promising "compliance with the AI standards" should therefore be able to say which standard, which version and which status.
Paragraph 4: errors, faults and feedback loops
Paragraph 4 requires the system to be as resilient as possible regarding errors, faults or inconsistencies within the system or its environment, in particular due to interaction with people or other systems. Technical and organisational measures are taken to that end. Robustness may be achieved through technical redundancy solutions, which may include backup or fail-safe plans. Recital 75 gives as an example a mechanism that lets the system safely interrupt its operation when certain anomalies occur or when it operates outside predetermined boundaries.
For the municipality this translates into two questions. What does the system do with an incomplete or unexpected application: does it give a recommendation anyway, or does it report that it cannot give a reliable one? And what does the department do if the system fails or is suspended: can it handle applications the old way, within the statutory deadlines? The second is not a product requirement, but it is a precondition for complying with Article 26(5). That provision says the municipality suspends use if it has reason to consider that use in accordance with the instructions may result in a risk within the meaning of Article 79(1). Suspending is not optional then. Without a fallback process, that obligation is hard to meet in practice.
The third subparagraph of paragraph 4 concerns systems that continue to learn after being placed on the market or put into service. These are developed in such a way as to eliminate or reduce as far as possible the risk of possibly biased outputs influencing input for future operations ("feedback loops"), and to ensure that any such loops are duly addressed with appropriate mitigation measures.
In the example the loop is obvious. The system flags applications for checks. Only flagged applications are checked, so errors are only found there. If the system learns from those findings, it learns that the group it already flagged is risky, and flags that group more often. Article 10(2)(f) already requires an examination of the data for possible biases, "especially where data outputs influence inputs for future operations" (Article 10). How to break that loop when choosing labels is covered in data quality and bias mitigation.
If the system keeps learning, that also raises the question whether it is still the same system. A system that has been substantially modified undergoes a new conformity assessment. For systems that continue to learn, changes that the provider pre-determined at the initial assessment and included in the technical documentation are not a substantial modification (Article 43(4)). So ask the vendor whether the system keeps learning, on which data, and which changes it has pre-determined.
Paragraph 5: attacks on the model
Paragraph 5 requires the system to be resilient against attempts by unauthorised third parties to alter its use, outputs or performance by exploiting system vulnerabilities. The technical solutions must be appropriate to the relevant circumstances and risks. For AI-specific vulnerabilities they include, where appropriate, measures to prevent, detect, respond to, resolve and control for attacks. Recital 76 adds that providers should take due account of the underlying ICT infrastructure.
| Attack (term in the Regulation) | What happens | What it looks like at the municipality (fictional example) |
|---|---|---|
| Data poisoning | Someone manipulates the training data | Manipulated files end up in a training set, so a certain pattern is never flagged again |
| Model poisoning | A pre-trained component used in training has been manipulated | The vendor builds on an external model containing a hidden behaviour |
| Adversarial examples or model evasion | Input designed to make the model err | An application is filled in so that the system wrongly sees it as low risk |
| Confidentiality attacks | Someone extracts information about the training data through the model | Clever queries to the system leak information about residents whose files were in the training data |
| Model flaws | Weaknesses in the model itself are exploited | A known blind spot is used systematically |
With the fourth category the AI Act touches the GDPR directly. In its Opinion 28/2024 on AI models (adopted on 17 December 2024), the EDPB states that AI models trained on personal data cannot, in all cases, be considered anonymous; this must be assessed case by case (point 34). As evidence of resistance, the EDPB mentions structured testing against attribute and membership inference, exfiltration, regurgitation of training data, model inversion and reconstruction attacks, noting that successful testing can only be evidence of resistance to the attacks tested (point 55). A provider that takes the fourth category of paragraph 5 seriously therefore also tests for what affects the GDPR.
If the system uses a language model that reads texts written by residents, hidden instructions in those texts also belong to the attacks paragraph 5 is about. How those work is covered in security risks of AI agents.
Article 42: when a certificate covers cybersecurity
Article 42 provides two routes by which a system is presumed to comply with the cybersecurity requirements of Article 15.
Paragraph 2: a certificate under the Cybersecurity Act. If a high-risk system has been certified, or a statement of conformity has been issued, under a cybersecurity scheme pursuant to Regulation (EU) 2019/881 whose references have been published in the Official Journal, it is presumed to comply with the cybersecurity requirements of Article 15, in so far as the certificate covers those requirements (Article 42(2)). That "in so far as" is the key: a certificate covering the security of a server or a software package says nothing yet about data poisoning or adversarial examples. So do not only ask whether there is a certificate, but also which requirements it covers.
Paragraph 3: the Cyber Resilience Act. Since the Digital Omnibus (Regulation (EU) 2026/1744), Article 42(3) states that a high-risk system falling within the scope of Regulation (EU) 2024/2847, the Cyber Resilience Act (CRA), and fulfilling the conditions of Article 12(1) CRA is deemed to comply with the cybersecurity requirements of Article 15. Article 12(1) CRA sets three conditions: the product fulfils the essential cybersecurity requirements for the product itself, the manufacturer's processes comply with the essential vulnerability handling requirements, and the EU declaration of conformity under the CRA demonstrates that the level of protection required by Article 15 is achieved. The conformity assessment then generally follows the AI Act procedure (Article 12(2) CRA). The CRA largely applies from 11 December 2027 (Article 71(2) CRA). Between 2 and 11 December 2027 the requirements of Article 15 therefore already apply to Annex III, while the CRA route is not yet available.
Two limits. First, both routes only cover cybersecurity. Accuracy and robustness remain outside; Article 12(1) CRA says so expressly. Second, the CRA applies to products with digital elements. According to recital 12 CRA, a cloud service is only covered when it is remote data processing for such a product; Software as a Service (SaaS) falls under the NIS2 Directive. A municipality that only uses the system as a cloud service therefore, in our reading, generally cannot rely on the CRA route. The requirements of Article 15 then apply in full, and the provider must demonstrate cybersecurity in another way.
How Article 15 connects with Articles 9, 10 and 26
Article 9: risk management. The provider sets up a risk management system that runs throughout the lifecycle (Article 9). It tests the system to identify the most appropriate measures and to ensure that it performs consistently for its intended purpose (paragraph 6), during development and in any event before placing on the market, against prior defined metrics and probabilistic thresholds appropriate to the intended purpose (paragraph 8). In choosing measures it gives due consideration to the technical knowledge, experience and training to be expected from the deployer and the presumable context of use (paragraph 5). In our reading, "appropriate" in Article 15 is therefore not a standalone number but the outcome of that risk management: which residual risk is acceptable determines which accuracy is needed. That matches Article 8(1): the requirements apply taking into account the intended purpose and the generally acknowledged state of the art, and the risk management system of Article 9 is taken into account when ensuring compliance.
Article 10: data. Training, validation and testing data sets are relevant, sufficiently representative and, to the best extent possible, free of errors and complete in view of the intended purpose (Article 10(3)), and take into account the specific geographical, contextual, functional or behavioural setting in which the system is used (paragraph 4). A model trained on data from large cities and used in a rural municipality may show different error rates from those promised in the instructions. A system trained and tested on data reflecting the setting in which it is used is presumed to comply with paragraph 4 (Article 42(1)).
Article 26: the municipality as deployer. The municipality takes appropriate technical and organisational measures to use the system in accordance with the instructions for use (Article 26(1)). To the extent it exercises control over the input data, it ensures they are relevant and sufficiently representative in view of the intended purpose (paragraph 4). It monitors the operation on the basis of the instructions for use and informs the provider where relevant (paragraph 5). This is where Article 15(3) meets Article 26: the accuracy figures in the instructions are the benchmark against which the municipality monitors. If actual error rates deviate clearly, that is information for the provider and possibly a reason to suspend use. If the municipality identifies a serious incident, it immediately informs first the provider and then the importer or distributor and the market surveillance authorities (paragraph 5; how that reporting chain works is set out in our piece on Article 73). The logs it needs for this are covered in logs of high-risk AI systems.
Human oversight depends on this too. The system must enable the people exercising oversight to properly understand its capacities and limitations and to remain aware of the tendency to automatically rely or over-rely on its output (Article 14(4), points (a) and (b)). A caseworker who does not know how often the system errs cannot do that.
The GDPR: accuracy, data protection by design and security
The AI Act leaves the GDPR unaffected (Article 2(7)). Three GDPR provisions run in parallel with Article 15, and they apply now, including to a system that only falls under the AI Act from 2027.
Accuracy (Article 5(1)(d) GDPR). Personal data must be accurate and, where necessary, kept up to date, and every reasonable step must be taken to ensure that inaccurate data are erased or rectified without delay (Article 5 GDPR). The EDPB points out that inaccurate data can lead to decisions being made on the wrong basis, including through automated decision-making or AI (Guidelines 4/2019, point 78). In our reading, for the municipality this means: a flag "additional check" that ends up in a file is information about that resident. If the flag turns out to be wrong, the file must show that.
Data protection by design (Article 25 GDPR). Both when determining the means of processing and during the processing itself, the municipality implements appropriate technical and organisational measures to implement the data protection principles, including accuracy, in an effective manner. Choosing a system with a particular error profile is such a determination of means. That makes the requirements it sets in the tender a GDPR question as well.
Security (Article 32 GDPR). The municipality and its processor implement measures to ensure a level of security appropriate to the risk, including as appropriate pseudonymisation and encryption, the ability to ensure the confidentiality, integrity, availability and resilience of systems, timely restoration after an incident, and a process for regularly testing and evaluating the measures (Article 32 GDPR). Paragraph 5 of Article 15 concerns the model; Article 32 GDPR concerns the entire processing, including integrations, storage and access.
In addition, a Dutch municipality is an essential entity by operation of law under Article 8(1)(h) of the Dutch Cybersecurity Act, which has applied since 15 August 2026. Its duty of care includes supply chain security (Article 21(3)(d) of the Dutch Cybersecurity Act). An AI system from a vendor is part of the network and information systems it must secure; see AI agents and the Dutch Cybersecurity Act.
What the municipality puts in its tender and contract
The municipality cannot deliver the requirements of Article 15 itself, but it can ask for them. A practical list:
- Accuracy in its own context. Ask for the metrics of Article 13(3)(b)(ii), and in any case ask for false positives and false negatives, per relevant group, on which test data and at which threshold. Ask how representative those test data are of your residents (Article 10(4)).
- A trial on its own data. Before deployment, run the system on historical applications that have already been decided and compare the outcome with the figures in the instructions. Record who does this and which legal basis the municipality has for it under the GDPR.
- Monitoring after deployment. Agree which performance data the vendor provides periodically, at which deviation it warns you, and which data you return to it for its own monitoring (Article 72(2)). Record whether that can be done in aggregated form.
- Continued learning and changes. Record whether the system keeps learning, on which data, which changes have been pre-determined (Article 43(4)) and how feedback loops are mitigated (Article 15(4)). Also agree that the vendor informs you in advance of a new model version and the figures that go with it.
- Fail-safe behaviour. Ask how the system responds to incomplete or unexpected input, and whether it can report that it cannot give a reliable recommendation.
- Cybersecurity of the model. Ask which tests the provider ran against the attacks in paragraph 5, and with what result. If you ask for a certificate, ask which requirements it covers (Article 42(2)) and whether the CRA route is available for this product (Article 42(3)).
- Supply chain arrangements. If the vendor builds on a model or component from a third party, ask how it secures its own obligations in that chain. The Regulation requires provider and supplier to conclude a written agreement for this (Article 25(4)).
- GDPR arrangements. Set out the security measures of Article 32 GDPR in the data processing agreement, and whether the vendor may use municipal data for training. If it may, the vendor is in our reading a controller for that purpose itself (Article 4(7) GDPR), and you need your own legal basis for that disclosure.
What the municipality arranges in its own process
- An owner of performance. Appoint someone who knows the figures in the instructions and periodically compares them with what the system does in practice.
- Signals from operations. Objections that are upheld, or flags that caseworkers often set aside, signal that accuracy in practice is lower than promised. Record how those signals reach the owner.
- A fallback process. Describe how the department handles applications when the system is suspended, and rehearse it.
- Training for caseworkers. Explain what the figures mean and when the system is less reliable, as part of human oversight (Article 26(2)).
- Input under control. Make sure the data the municipality supplies itself are relevant and sufficiently representative in view of the intended purpose (Article 26(4)); completeness and being up to date also follow from Article 5(1)(d) GDPR.
- DPIA and fundamental rights impact assessment. Use the information from the instructions for the DPIA (Article 26(9) AI Act and Article 35 GDPR). As a body governed by public law, the municipality also carries out a fundamental rights impact assessment before deployment (Article 27); see the FRIA template.
From when?
Article 15 is in Chapter III, Section 2. For high-risk systems in Annex III it applies from 2 December 2027, for systems under Annex I from 2 August 2028 (Article 113, third paragraph, point (c), as amended by Regulation (EU) 2026/1744). A system that assesses whether residents are eligible for an essential public assistance benefit, or helps to reduce, revoke or reclaim it, falls in our reading under Annex III, point 5(a). If it performs profiling of natural persons, it is always high-risk (Article 6(3), last subparagraph).
If the system was placed on the market or put into service before 2 December 2027, the high-risk obligations only apply if it undergoes significant changes in its design from that date. Providers and deployers of systems intended to be used by public authorities must in any event comply by 2 August 2030 (Article 111(2)). For a tender running now, that is no reason to leave Article 15 out of the specifications: the system you buy now must comply later, and the GDPR and the Dutch Cybersecurity Act already apply.
Example 2: the bank and the credit assessment
A bank using AI to evaluate the creditworthiness of individuals or establish their credit score uses a high-risk system under Annex III, point 5(b) (except AI used to detect financial fraud). Banks often develop such models themselves. The bank is then the provider and Article 15 rests on it directly: it determines the appropriate level, declares the metrics in the instructions for its own staff, and tests against attacks.
Three points differ for the bank compared with the municipality.
Accuracy has two sides. A model that is too strict wrongly rejects people; a model that is too lenient grants credit that cannot be repaid. Which side weighs more belongs in the risk management of Article 9, and differences between groups belong in the bias examination of Article 10(2)(f).
Monitoring runs through governance. For a deployer that is a financial institution subject to governance requirements under Union financial services law, the monitoring obligation of Article 26(5) is deemed fulfilled by complying with those governance rules (Article 26(5), second subparagraph). In our reading, that does not remove the obligation to inform the provider and the authority of a risk or a serious incident; the safe course is to set up that reporting route separately.
DORA for the ICT side. A bank falls under DORA (Regulation (EU) 2022/2554), applicable since 17 January 2025, with its own ICT risk management framework (Chapter II, Article 6 DORA). That framework covers all information and ICT assets, including software, and so the AI system itself. Article 15(5) adds the AI-specific attacks as an explicit design requirement for the provider.
In addition, as a deployer under point 5(b), the bank carries out a fundamental rights impact assessment (Article 27(1)). How a credit decision stays explainable is covered in from scoring algorithm to transparent credit decision.
What to do now
- Establish whether the system is high-risk and who the provider is. If you build it yourself, put your name on it or substantially modify it, you are the provider and Article 15 rests on you (Article 25(1)).
- Ask for the figures. Ask for accuracy, robustness and cybersecurity as Article 13(3)(b)(ii) describes them, broken down and with the circumstances in which they do not hold.
- Test them against your own population. A trial on your own, already decided cases says more than a number in a bid.
- Put the arrangements in the contract. Monitoring, continued learning, new versions, fail-safe behaviour, tests against attacks, certificates and what they cover.
- Set up your own process. An owner, signals from objections and operations, a fallback process, training and a DPIA.
How the GDPR and the AI Act interact more broadly is set out on the page on the GDPR alongside the AI Act. Unsure whether your system is high-risk and which obligations then apply? With the case check you describe the application in a few questions and see which AI Act and GDPR obligations come into view.
Frequently asked questions about Article 15 AI Act
Sources
Newsletter
Every Tuesday, the AI Act week ahead in 5 minutes
A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.
Practical and short · No spam · One-click unsubscribe